OpenAI Has Developed an Advanced Watermarking Tool for ChatGPT—but Will It Be Released?

OpenAI has been developing a system for watermarking text generated by ChatGPT, a feature that has been ready for about a year, according to a report by The Wall Street Journal.

This watermarking technology adjusts how the AI predicts words and phrases, creating a detectable pattern in the text. While the company has been internally divided over the release of this tool, it has not yet made it publicly available.

The Rise of ChatGPT and the Cheating Problem

ChatGPT's ability to generate human-quality text on any topic raised concerns about academic dishonesty. Students could potentially use it to churn out essays without any effort. OpenAI, however, lacked any detection method at the time.

Development of the Watermark Tool

Over the past two years, OpenAI has developed a "text watermarking" tool for ChatGPT. This tool, according to reports, boasts a remarkable 99.9% accuracy in identifying ChatGPT-generated content. It achieves this by subtly manipulating the AI's selection of words within a sentence, leaving behind a watermark invisible to the human eye.

Public Support for Detection, User Resistance to Watermarking

OpenAI commissioned a survey in April 2023 that revealed strong global support for a tool that could detect ChatGPT text. However, a different survey conducted in the same month showed that nearly 30% of ChatGPT users threatened to abandon the platform if watermarked content became the norm. This user resistance presented a significant hurdle for OpenAI.

Beyond User Base Concerns

OpenAI also considered the impact on non-native English speakers who might rely on ChatGPT for writing assistance. Additionally, the potential for "bad actors" to circumvent the watermarking system through techniques like rewording or translation raised concerns about the tool's long-term effectiveness.

Alternatives and the Future of Detection

OpenAI is exploring alternative methods like embedding metadata into the text itself. This approach, unlike watermarks, offers cryptographic verification, eliminating false positives. However, it remains unclear when, or even if, this alternative will be implemented.

A Wider Issue of AI Text Attribution

OpenAI isn't alone in this challenge. Google, for instance, also has a text watermarking tool for its AI model, Gemini. Similar concerns remain around user backlash and circumvention methods. Both companies are actively developing methods to identify AI-generated visual content as well. Ideally, the industry will see a standardized approach to watermarking AI-generated text in the near future.

Via

Source